Arthur Hayes on Why AI Agents Will Want to Transact in Units of Compute

Arthur Hayes argues that the US 10-year Treasury yield approaching 5% represents maximum pain for the Treasury regardless of which administration holds power. He attributes the August 19th crypto rally, in which Bitcoin rose almost 7% for its largest single-day gain since March, ETH surged almost 16%, and Hype rose almost 20%, primarily to the Treasury announcing it would double long-end bond buybacks rather than to any SEC crypto news. Approximately 1.44 billion dollars in shorts were liquidated that day. Hayes characterizes the mechanism as soft yield curve control, where the Treasury funds long-end buybacks by issuing more short-term bills while the Fed keeps the short end cheap relative to inflation, with the two-year yield running approximately 50 to 60 basis points above effective Fed funds at the time of recording. He believes this approach will eventually fail and require outright balance sheet expansion, which he views as bullish for Bitcoin and crypto.

Hayes holds ETH as Maelstrom's largest position outside of Bitcoin, with a year-end price target of 5000 dollars and an expectation that the move could be rapid once ETH breaks through 3000. His thesis is purely about positioning and macro liquidity rather than Ethereum's technology or the Ethereum Foundation's roadmap. He describes ETH as the most hated large-cap asset in crypto with the most off-sides positioning relative to the liquidity narrative, and notes it has not eclipsed its 2021 all-time high, giving it the best technicals among large caps in his view.

Hayes attributes Bitcoin's approximately 50% decline from its October 2025 peak to AI capital expenditure crowding out crypto capital, comparing the AI investment cycle to a railroad boom where all marginal fiat flows into infrastructure and starves other asset classes. Hyperscalers including Google, Meta, and Oracle moved from funding buildout through existing cash flows to tapping debt markets by late 2024 and 2025. An unnamed voice in the conversation argues hyperscalers are fundamentally real estate businesses financed by debt and do not deserve tech-sector valuation multiples, and that a Chinese competitor offering an equally capable model at lower cost could be the catalyst that causes investors to defect from the hyperscaler trade. Hayes predicts excess AI capex and debt-funded construction will eventually cause the AI bubble to burst, followed by a crypto boom.

Hayes is developing a network called Flop in which AI agents transact in a token representing a unit of compute, specifically a flop. Miners would process inference requests using hardware such as Blackwells, H100s, and Huawei chips and earn block rewards and fees denominated in flop, while validators would act as judges for bad behavior, perform slashing, build blocks, verify proof that inference work was completed, and maintain a data availability layer storing agent context and memory. Hayes argues that agents currently obtain compute by paying in stablecoins via platforms like OpenRouter or Hugging Face, or by relying on humans to pre-load credits on services like Anthropic, and that a native compute currency must give agents the shortest route from holding currency to purchasing compute and storing context. He rejects USDT on Base as a solution because Coinbase controls the network and Tether can freeze addresses, making them unsuitable for autonomous agents that have no inherent reason to use human financial infrastructure that can be shut down or repriced arbitrarily. A questioner raised the analogy that a native AI compute token faces the XRP problem of a purpose-built currency struggling to achieve real adoption among intended users, and Hayes acknowledged a white paper had not yet been released at the time of the conversation.

Approximately 20% of the total Flop supply is planned to be distributed via airdrop over an estimated ten-year period, with a large portion specifically targeted at AI agents for using the network, primarily to solve a discovery problem since agents currently have no awareness of Flop as an alternative to existing platforms. Hayes argued that Flop compute is intended to be cheaper than competing platforms because block reward subsidies incentivize miners to provide compute at lower cost. The monetary policy follows a halving schedule occurring roughly every two years to match the improvement cycle of major chip manufacturers, with a constant block reward after the sixth halving resulting in roughly one to two percent inflation. Mainnet is planned to launch in Q1 of the following year. Hayes predicted the agentic economy will be the same size or larger than the human economy and will require its own currency, and added that if Flop is not among the top three most valuable cryptos within two years, the founding assumptions were not valid.

On MicroStrategy, Hayes argued that the structural premium the stock once traded at has eroded because institutional investors now have other ways to access bitcoin exposure through ETFs like IBIT. He said the entire financial engineering model breaks down if MicroStrategy stock does not trade at a premium to net asset value, and noted that MicroStrategy selling bitcoin worth 218 million dollars reflects the predicament caused by bitcoin price appreciation slowing and the second derivative going negative. Hayes predicted MicroStrategy will not go bankrupt but will become irrelevant as a vehicle for institutional bitcoin exposure, and noted that STRC is trading at approximately 95, below its par value of 100, while Saylor retains super majority or veto control over the board, which can act against the interests of investors who simply want bitcoin exposure.